01 / 03
Theory of Agency in AI Interaction (TAAI)
This work formulates four fundamental laws that define the boundaries of human interaction with artificial intelligence, the conditions for preserving human agency, and the distribution of responsibility.
A detailed article on this topic will be published in the near future. Below is a brief summary of the theory.
Problems with AI Systems
Increasingly, humans are delegating to AI not only information retrieval, but also planning, communication, decision-making, and the execution of actions. In this context, the question of where human intent ends, where the system’s execution begins, and who bears responsibility for the consequences becomes fundamental.
Problems that arise from ignoring human agency:
- The LLM misinterprets the task.
- Errors and the accumulation of errors in chains. Autonomous AI agents.
- The wrapper extends beyond the scope of the assignment.
- A person accepts an erroneous result.
- AI poses as a human.
- AI-generated content is indistinguishable from human-created content.
- AI simulates a friend or a psychologist.
- The user and the developer shift responsibility onto each other.
- The developer masks flaws in the shell by attributing them to the unreliability of the LLM.
Laws of Agency in AI Interaction
But to what extent does human agency manifest itself within the contemporary socio-economic model of the digital environment?
- The Law of Agency: An individual must personally define their own goals, intentions, and the limits of their authority, and must bear the consequences of their decisions.
- The Law of Uncertainty: LLM output cannot be considered guaranteed to be accurate and requires verification.
- The Law of Transparency: A person must be aware that they are interacting with AI-generated content.
- The Law of Responsibility: If the developer of an AI system complies with the first three laws, then the user bears responsibility for the use of the AI system and for the consequences of any actions taken based on its output.